The Centaur Era: Half Human, Half Machine

In June 1998, a year after IBM’s Deep Blue defeated him in a match the world treated as a referendum on the human mind, Garry Kasparov sat down in León, Spain, not to play against a machine but to play with one. Across the board was Veselin Topalov. Each man had a PC and a chess engine. Kasparov called the format Advanced Chess. The idea was architectural. If a machine could search more positions than any grandmaster could hold in working memory, and a human could still decide which lines were worth searching, then the interesting unit of intelligence was no longer either participant. It was the coupled system.

A month earlier, in ordinary rapid chess, Kasparov had beaten Topalov 4–0. In León they drew 3–3. Kasparov’s own account, written a decade later, is that his advantage in calculating tactics had been nullified by the machine. The engines did not automatically produce a higher form of play. They levelled the board. Whatever the coupled system was going to add would have to come from somewhere else. Collaboration was itself a skill.

That is the part of the story worth keeping. Chess is the metaphor, not the subject.

Process, then capability

The result that made the metaphor travel arrived in 2005, in an online “freestyle” tournament that permitted any combination of humans and computers. Grandmasters entered with strong hardware and strong colleagues. So did Hydra, a chess-specific supercomputer that, the same month, beat one of the ten highest-rated players in the world by five and a half games to a half. Neither version of Hydra reached the quarterfinals. The winners were two American amateurs, Steven Cramton and Zackary Stephen, rated well below master level, operating three ordinary PCs. Kasparov’s distillation, written five years later, has been repeated ever since: weak human plus machine plus better process was superior to a strong computer alone, and, more remarkably, superior to a strong human plus machine plus inferior process.

What the amateurs were good at was not chess in the romantic sense. They ran four engines across their three machines, and in Kasparov’s telling their skill lay in manipulating and coaching those engines to look very deeply into the positions that mattered. The machine was not an oracle. It was an instrument whose attention could be aimed. The human contribution was already migrating: away from making the move, toward deciding how the move would be produced.

That migration is the subject here. Not whether AI will “augment” us. It already does. The question is what happens when the human contribution shifts from performing the work to deciding what work should happen, what constraints should govern it, when the machine has done enough, and when control must return to a person.

Five stages of a coupled system

The shift does not arrive all at once. It proceeds through stages that look, from the inside, like ordinary improvements in tooling.

First, a human performs the work. Judgment, memory, drafting, calculation, and execution live in one skull. Errors are human errors. Responsibility is easy to locate because almost nothing else was involved.

Second, a human uses a machine as a tool. Spreadsheet, compiler, search engine, IDE. The human still frames the problem, sequences the steps, and recognizes when the output is wrong. The machine compresses labor. It does not yet propose the labor.

Third, a human delegates substantial cognitive work to an AI system. Draft this. Compare these options. Write the function. Summarize the docket. Simulate the scenario. The human remains, in principle, the author. In practice the human becomes an editor of machine-generated intermediates. Much of what used to count as thinking now arrives pre-shaped.

Fourth, a human designs, supervises, and governs systems of agents that perform work. The object of attention is no longer a document or a function. It is a process that can continue while the human is elsewhere. The human sets objectives, boundaries, budgets, tools, and stopping conditions. Execution becomes a property of the system. The human role becomes gubernatorial.

Fifth, still partly prospective but already visible in narrow domains, the machine becomes capable enough that the human’s direct contribution is no longer justified by performance alone. Intervention starts to look like noise. The remaining arguments for keeping a person in the circuit are no longer about quality of output. They are about legitimacy, liability, values, and the right to refuse a course of action that would, by the machine’s lights, succeed.

Most knowledge work is now distributed across stages two through four, often in the same afternoon. That is why the old slogan, humans and AI work well together, explains almost nothing. It does not tell you which stage you are in. It does not tell you what the human is still for.

What the machine takes, and what does not automatically remain

It is tempting to draw two columns and call the problem solved. On one side: search, recall, calculation, drafting, coding, comparison, simulation, classification, monitoring, execution, coordination among machines. On the other: intent, values, contextual judgment, constraints, skepticism, escalation, acceptance criteria, accountability, the decision to intervene, the decision that the system is solving the wrong problem.

The columns are not false. They are unstable.

They assume complementarity that does not have to persist. Chess already showed this. Centaur play was interesting while humans and engines had different failure modes. Humans blundered tactically; engines were strategically crude, or at least differently crude. A person who could aim the search still added something. Even then the margin was thinner than the story suggests. A later engine-based analysis of the freestyle games from 2005 to 2008 found that the human–machine teams matched a strong engine’s preferred move more often than computers playing alone did, but did not clearly beat those computers on average error. Then the engines became overwhelmingly strong, first by raw search and later by self-play learning that produced superhuman evaluation from nothing but the rules of the game. The value of overriding the machine declined. In many positions the human’s distinctive contribution became the ability to make the line worse. Complementarity had been a temporary fact about two kinds of error, not a law of nature.

The pattern is not confined to the board. A 2024 meta-analysis of 106 experiments found that, on average, human–AI combinations performed worse than the better of the human or the AI alone. The losses were concentrated in decision tasks, and they appeared precisely where the AI alone had outperformed the human. Those were laboratory tasks, and they say nothing about governance. They do say that complementarity is not something you get for free by adding a person.

Knowledge work is not chess. The board is not fully observable, the rules are not closed, and winning is not checkmate. The analogy is useful only this far: complementarity can evaporate inside a domain without the domain ceasing to matter. It does not follow that every domain will take the same curve on the same timetable. It does follow that we should not build a theory of human authority on the hope that machines will remain blind where we are sighted.

Shared control is not a feeling

The centaur, properly understood, is not a mascot for teamwork. It is an architecture of shared control.

The important questions in that architecture are operational. Who may start a process, extend it, spend money, change state in the world, send a message that cannot be unsent? What evidence is required before the next step is authorized? What conditions force a halt even if the system can continue? Who is answerable when the halt does not come?

These are not the same question as “is a human involved?” A human can be involved in ways that do not constitute control.

Consider the phrase human in the loop. It sounds like a safeguard. Often it is a workflow. An agent produces a recommendation. A person clicks approve. Sometimes that click is governance. Often it is a cost-saving ritual. Checking the work is more expensive than generating it. The reviewer cannot independently reproduce the reasoning. The alternatives were ranked before the person arrived. The default is already selected. Under time pressure, authorization substitutes for evaluation.

At that point the human is not supervising the system so much as notarizing it. The loop still contains a person. The person is no longer doing the thing the loop was invented to preserve.

The problem sharpens when the machine becomes better at recognizing errors than the supervisor. A reviewer of ordinary tools can catch a bad formula because they still know how the formula should behave. A reviewer of multi-agent output is often looking at a finished artifact whose production path they did not walk and cannot cheaply reconstruct. If the system is also better at spotting inconsistencies in its own draft, the reviewer’s remaining advantage is not detection. It is refusal: the capacity to say that a well-formed answer is an answer to the wrong question, or that a locally optimal action violates a constraint the system was never told to treat as hard.

That contribution is real and fragile. It depends on still being able to see the frame rather than only the product. The more fluent the system becomes at producing work that looks as if it came from a sound frame, the harder that seeing gets.

Operator to governor

Increasing autonomy changes the human role from operator to governor. The sentence is easy to accept and hard to inhabit.

An operator is coupled to the work in time. The operator feels the resistance of the material. Errors arrive while there is still a chance to reverse them. Skill is maintained by use.

A governor is coupled to the work through policy. The governor defines an envelope: purpose, permitted tools, data the system may see, actions it may take, thresholds at which it must stop and ask, conditions under which it may not continue even if it can. The governor samples, investigates exceptions, and changes the envelope when the world changes.

The skill is more abstract, and it fails differently. Operators fail by doing the wrong thing. Governors fail by authorizing the wrong regime of action, or by failing to notice that the regime has drifted. A system can be capable of continuing long after it has ceased to be authorized to continue. If no one specifies the difference, capability will be treated as permission. That is how problems get moved rather than solved. The visible task completes. The risk, the ambiguity, and the accountability migrate to whoever still holds the approval button.

The previous essay in this sequence asked which judgments belong in durable structure and which must stay live, and warned against writing an open question in the file format of policy. Bounded autonomy is the same question asked about action rather than about records. The system may act without asking, but only inside constraints that are actual constraints, not guidelines the model is politely asked to consider. Escalation is not a confession of failure. It is the designed response to leaving the envelope. Stopping conditions belong in the work specification, equal in status to the objective. A system that can explain why it is still inside the envelope is more governable than a system that can only present a polished result.

None of this is guaranteed by placing a person somewhere in the diagram. Governance is a practice. It atrophies when the governor no longer knows which questions would reveal that the envelope has been left. It also atrophies when every action requires a click, because attention is finite and ritual approval trains an institution to stop looking.

Performance is not legitimacy

There is an uncomfortable possibility that the centaur era is transitional.

Centaur chess was genuine for a window of years in which humans and machines had complementary strengths and the interface still rewarded craft. Then the machines pulled away. Kasparov, looking back over the whole arc of human–machine chess, put the general shape plainly: the computers went quickly from too weak to too strong, and the genuinely fascinating contests occupied a span of about ten years in between. People still play chess. They play it in a world where the strongest play is not human, and where a person at the keyboard is no longer expected to improve on the strongest engine’s move.

Other cognitive domains may not close that cleanly. Law, medicine, management, engineering, and politics are soaked in contested purposes. There is no single evaluation function hiding under the paperwork. Even so, inside many subtasks, such as retrieval, first drafts, code synthesis, reconciliation, and log monitoring, the performance gap is already large. The human who insists on performing those subtasks by hand is not defending judgment. They are defending a habit.

If that trajectory continues, the centaur era is the period in which society still gets to decide which forms of human authority are worth preserving before capability alone answers the question for us.

That distinction is the one most easily lost. A machine being better at making a decision does not establish that the machine ought to have authority to make it. Performance and legitimacy are different variables. Performance asks which arrangement produces the better result on a specified measure. Legitimacy asks who is entitled to bind others, spend a common resource, impose a risk, close off a future, or declare that this is the problem being solved.

Some decisions should remain human-controlled even when a system would score higher on the local metric. Not because human judgment is superior, since often it is not, but because some acts are exercises of authority rather than of optimization. The measure itself may be in dispute. People may refuse to be governed by a process they cannot contest. Accountability that cannot land on a person becomes, in practice, accountability that lands on no one.

The reverse error is just as available: treating the final button-press as proof that authority has been retained. A button is not a philosophy of control. If the person pressing it cannot evaluate the action, cannot refuse it without institutional penalty, and cannot explain it afterward except by pointing at the model, then the human in the loop is a liability shield. That is not preservation of agency. It is the appearance of agency purchased at the price of its contents.

What we are still deciding

The important question is no longer merely what the machine can do. It is what role remains for the human when the machine can do it better. And it is which of those remaining roles we are prepared to defend on grounds other than accuracy.

That is not a question a benchmark can close. It has to be answered domain by domain, with an unsentimental inventory of what the human is actually contributing: not what we hope they contribute, not what the org chart says, but what would change if they stopped. In some places the honest answer will be almost nothing about the quality of the artifact, and a great deal about who may be blamed. That answer is information. It tells you the human role has already become ceremonial, and that ceremony is being asked to carry a weight it cannot bear.

In other places the human is still doing the only work that makes the rest of the work make sense: choosing the objective, naming the constraint that must not be optimized away, noticing that the system is succeeding at a proxy, deciding that continuation is no longer authorized. Those are governor functions. They survive the loss of complementarity only if they are treated as real work—trained, resourced, and allowed to halt a process that is performing well.

The centaur was never a stable creature. In the myth it is powerful because two natures share one body, and dangerous for the same reason. The era named after it will not last because the image appeals. It will last as long as the human half can still change what the system does, for reasons we can state, at moments that still matter. After that we will not be centaurs. We will be the people who did or did not decide, while the decision was still ours, which kinds of authority a better machine is not allowed to inherit.

Provenance

Abstract

Advanced Chess in 1998 and freestyle chess in 2005 are the standard evidence that a human and a machine, coupled well, outperform either alone. This essay takes the less-quoted half of that history seriously: the coupling was a skill, its advantage was thinner than the slogan suggests, and it did not survive the machines' improvement. It proposes a five-stage progression of human–machine work, from performing the task to governing systems of agents, and argues that the familiar two-column division of labour assumes a complementarity that does not have to persist. What remains for the human is then reframed as an architecture of shared control: who may start, extend, spend, and halt, on what evidence, and who answers when the halt does not come. The operator-to-governor transition, bounded autonomy, and stopping conditions are the practical form of that architecture; ceremonial approval is its failure mode. The closing distinction is between performance and legitimacy. A machine being better at a decision does not establish its authority to make it, and the centaur era is the interval in which that authority is still being allocated deliberately rather than by capability alone.

Sources (9)

Inspection scope: presence here does not imply full-text access to every cited work or reproduction of any result. Links were inspected during review of this essay on September 22, 2026. Where a secondary reference (an encyclopedia entry) is listed, it was used for dates and match details and is marked as secondary; primary contemporaneous reports were preferred wherever they were available.

Fact-check table (27)

VERIFIED means the specific claim is supported by the identified inspected material. QUALIFIED marks an interpretation, a secondary source, or a result whose scope matters. CORRECTED marks a claim changed from the draft after verification. AUTHOR OBSERVATION and ORIGINAL SYNTHESIS are labeled as such.

Editorial note: original synthesis

Inherited. The chess history is Kasparov's, in the 2010 essay, with the contemporaneous ChessBase reports supplying names, ratings, dates, and results. The finding that human–AI combinations on average underperform the better party alone, with losses in decision tasks, is Vaccaro, Almaatouq, and Malone's, and was already used in What Should the Agent Have to Figure Out? for a narrower purpose. The observation that self-play learning produced superhuman chess evaluation from the rules alone is Silver et al.'s. The thinness of the freestyle margin is Regan's informal finding. The distinction between what has been settled by someone with standing and what must remain live is inherited from Part 3 of this series and is invoked here by link rather than restated.

Adapted. Kasparov's "too weak to too strong" arc is used as a template for how complementarity can close inside a domain; he wrote about chess, not about knowledge work, and the essay says so. Vaccaro et al. is applied to the claim that complementarity is not automatic; the meta-analysis concerns laboratory task performance and is explicitly disclaimed as evidence about governance.

Analogical only. Chess throughout, with the limits stated in the body: the board is fully observable, the rules closed, and the objective singular, none of which holds for the domains the essay is about. The centaur of myth.

This essay's contribution. The five-stage progression of human–machine work; the claim that the standard two-column division of labour assumes a complementarity that does not have to persist; the centaur as an architecture of shared control rather than a metaphor for teamwork; the reading of "human in the loop" as often a workflow in which authorization substitutes for evaluation, with the reviewer's durable advantage being refusal rather than detection; the operator-to-governor transition and the observation that capability is treated as permission when nobody specifies the difference; bounded autonomy with stopping conditions as first-class parts of a work specification; the distinction between performance and legitimacy as separate variables; and the proposal that the centaur era is transitional and is the interval in which the allocation of authority is still being decided deliberately.

Corrected after review. The names of the 2005 winners. The claim that Kasparov said both players had failed to combine human and machine skill, replaced with his actual statement that the machine nullified his tactical advantage. The claim that the winners worked by noticing disagreement between engines, replaced with Kasparov's description of their coaching the engines to search deeply. The four-phase pattern attributed to Kasparov, replaced with his "too weak to too strong" and "span of ten years." "Imaging reads" removed from the list of subtasks. The assertion that the strongest chess is not a hybrid was softened to an expectation rather than a measurement. A sentence on the thinness of the freestyle margin and a paragraph on the Vaccaro meta-analysis were added to the section on unstable complementarity, because the argument that complementarity is temporary needed evidence beyond one tournament.

Limits. The chess history rests on Kasparov's own retrospective account and two contemporaneous news reports, plus encyclopedia entries for dates. No tournament archive or game record was inspected. Regan's study is an informal web analysis, not a peer-reviewed result. The essay's central claims about governance, legitimacy, and the ceremonial character of much human-in-the-loop review are arguments, not findings; no organization's approval process was studied for this essay. Unlike Parts 1 through 3, this essay contains no first-person repository case, and it should be read as the series' synthesis rather than as evidence.